164 - Feature Interaction-Informed Deep Radiomics for Prediction of Distant Metastasis-Free Survival in Head and Neck Cancer

08:50am - 09:00am ET

Room 156

Presenter(s)

Andrew Heider, BS - Stanford University, San Jose, CA

R. Hou^1,2, Y. Shen¹, C. Zhang², M. T. Islam², X. Fu¹, L. Xing², and A. Heider²; ¹Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ²Department of Radiation Oncology, Stanford University, Stanford, CA

Purpose/Objective(s):

Distant metastasis remains a dominant failure pattern after definitive therapy for head and neck (HN) cancer, yet standard clinicopathologic factors insufficiently stratify metastatic risk. Conventional radiomics treats imaging features as independent variables (“bag-of-words”), forcing deep models to infer feature relationships from scratch. We propose an interaction-informed deep radiomics framework that explicitly models feature-feature interactions from pretreatment CT to predict distant metastasis-free survival (DMFS) and uncover microenvironmental programs linked to metastatic propensity across multi-institutional cohorts.

Materials/Methods:

This large-scale study integrated 3,421 HN patients from four independent cohorts (RADCURE, HN1, HN-PET-CT, TCGA-HNSC). We utilized the OmicsMap framework to transform high-dimensional radiomic features into 2D topological maps for deep survival modeling. Prognostic discrimination was assessed using Harrell’s C-index and Kaplan-Meier analysis. Model interpretability was evaluated using SHAP attribution and visualization of risk-associated OmicsMap patterns. For biological validation, patients with matched CT and RNA-seq data (n=94) in TCGA-HNSC cohort underwent differential expression analysis and Gene Set Enrichment Analysis to profile tumor microenvironment and immunotherapy-relevant signatures.

Results:

The model demonstrated consistent prognostic performance with C-indices of 0.742 (RADCURE), 0.768 (HN1), and 0.671 (HN-PET-CT), and significantly separated DMFS risk groups across cohorts (P<0.01). Integrating key clinical factors further improved external generalizability (C-index 0.864 in HN1; 0.730 in HN-PET-CT). SHAP analyses showed that predictions were dominated by wavelet-transformed texture heterogeneity features from GLSZM, GLCM, and GLDM families, reflecting gray-level nonuniformity, local homogeneity, and entropy-related disorder. OmicsMap revealed distinct risk patterns: high-risk tumors exhibited compact, high-activation regional clusters, whereas low-risk tumors displayed diffuse, low-intensity spatial distributions. Radiogenomic profiling supported biological coherence: imaging-defined high-risk tumors aligned with fibrosis-prone, immune-excluded programs with proliferative/hypoxic enrichment, whereas low-risk tumors showed immune-active biology characterized by interferon-a signaling and increased eosinophil-associated signatures.

Conclusion:

Pretreatment CT captures an interpretable and biologically grounded radiomics phenotype that generalizes across institutions for DMFS prediction in HN cancer. This framework may enable risk-adapted surveillance, trial enrichment, and biomarker-driven post-definitive intensification strategies.